CienaLLM

CienaLLM (Climate Impact Extraction from News Articles using LLMs) is a modular Python framework for schema-guided Generative Information Extraction (GenIE) from climate-related news using open-weight Large Language Models (LLMs). It enables configurable prompts, multi-step extraction pipelines (summarization, self-criticism, chain-of-thought), and structured JSON outputs.

The framework was developed and evaluated in Vela-Tambo et al. (2025, to be published), focusing on extracting drought impacts from Spanish news articles. Its modular design makes it easily extensible to other hazards, domains, and languages.

---

Features

• Schema-guided extraction with configurable prompts and schemas
• Multi-step pipelines: summarization, chain-of-thought, self-criticism
• Structured output parsing and validation with Pydantic
• LLM orchestration through LangChain and Ollama
• Modular design with reproducible experiments via YAML configuration
• Support for both local inference and cloud-based LLMs

---

Project Structure

.
├── README.md                       # This file
├── LICENSE                         # MIT license text
├── pyproject.toml                  # Poetry configuration file
├── poetry.lock
├── src                             # Source code
│   └── ciena_llm
│       ├── article                 # Article processing and handling
│       ├── llm                     # LLM interaction
│       ├── chain                   # Each LLM step definition
│       ├── extraction_schema       # Extraction schema definition
│       ├── prompt                  # Prompt definitions and management
│       ├── output                  # Output handling and formatting
│       ├── config                  # Configuration handling
│       │   └── config.yaml         # Default configuration file
│       └── __init__.py
├── experiments                     # Scripts for experiment execution
└── tests                           # Scripts for testing
    ├── common
    ├── test_event.py               # Test script for event relevance extraction
    ├── test_impact.py              # Test script for impact extraction
    ├── test_location.py            # Test script for location extraction
    └── test_hail_event.py          # Test script for hail event extraction

---

Installation

Ensure you have Python 3.10+ installed. Install dependencies with Poetry:

poetry install

Alternatively, you can install dependencies using another method (e.g., pip) if you prefer.

---

Usage

1. Configuration: Create a script that uses the ciena_llm package. Example scripts can be found in the tests directory.

   You can either:

   • Set up your configuration in config.yaml, or

   • Pass a config file directly in your script.

2. Execution: Run your script with Poetry:

   poetry run python your_script.py

---

Datasets

This repository does not include datasets. The experiments described in the accompanying paper used re-annotated drought-related Spanish news datasets. These will be released separately upon publication.

You can run the framework on your own corpora of news articles formatted according to the NewsArticle schema.

---

Citation

If you use this framework in your research, please cite:

@misc{vela2025cienallm,
  author = {Javier Vela-Tambo and Jorge Gracia and Fernando Domínguez-Castro},
  title = {CienaLLM: Generative Climate-Impact Extraction from News Articles with Autoregressive LLMs},
  year = {2025},
  note = {To be published}
}

---

Contributing

Contributions are welcome! Please open an issue or pull request if you’d like to improve the code, add features, or report bugs.

---

License

This project is licensed under the MIT License – see the LICENSE file for details.